PaperlayerView exact revision on arXiv

Source-linked AI summary

Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception

Lai Wei, Yatai Li, Liangbo He, Jun Lan, Renjie Gu, Linhan Cao, Junkai Lin, Huijia Zhu, Weiqiang Wang, Linghe Kong, Yue Wang, Zhuosheng Zhang, Weiran Huang

arXiv:2610.37037v1cs.CVcs.AIcs.CLcs.LG

TL;DR

Spatial perception and multimodal long-context understanding depend on connecting visual evidence, but high-quality training data for these tasks are difficult and costly to construct. EvidenceConnect synthesizes and verifies single-image, multi-region questions, and RLVR-trained ZwZ2 improves across benchmarks and downstream task types. The results suggest that evidence connection is a transferable training target.

  • Problem

    High-quality training data for spatial perception and multimodal long-context understanding are difficult and costly to construct, despite their importance to embodied AI and multimodal agents.

    Source: §1 · p. 2

  • Method

    EvidenceConnect synthesizes single-image questions requiring comparison, aggregation, or chaining across regions, verifies answers and evidence dependence, and uses the data for RLVR training.

    Source: §1 · pp. 2–3, §3 · p. 4

  • Results

    ZwZ2 improves over its base models on all 25 benchmarks at 4B, 9B, and 27B, with average gains of 5.39, 4.96, and 4.99 points, respectively.

    Source: §4.2 · p. 8, §5 · p. 10

  • Takeaways & Limitations

    The results suggest that connecting evidence is a transferable training target, including from single-image training to long-context understanding and multimodal agent tasks.

    Source: §5 · p. 10

    Because training examples contain a single image and short context, gains narrow at the longest contexts, where retrieval likely becomes the main difficulty.

    Source: §4.5 · pp. 9–10

Abstract

from arXiv · show

Multimodal perception forms the foundation of multimodal intelligence. In particular, spatial perception and multimodal long-context understanding are two pivotal tasks that underpin embodied AI and multimodal agents. However, synthesizing targeted, high-quality training data for these tasks remains difficult and costly. Our key insight is that these tasks share an underlying capability, which we call connecting evidence: identifying relevant visual facts, binding them to the correct entities, recovering their relations, and composing them into an answer. Crucially, targeted training data are far easier to synthesize for this capability than for these tasks themselves. To strengthen this capability, we propose EvidenceConnect, an easy-to-implement data synthesis framework that uses zoomed views of related regions within a single image to generate questions requiring comparison, aggregation, or chaining of visual evidence. To ensure quality, a reference answer is retained only when independent VLMs agree on it, and region-ablation tests discard questions whose answer survives the removal of selected regions. We then train base VLMs on the resulting data with RLVR, obtaining Zooming without Zooming v2 (ZwZ2). Although the synthesized data are not tailored to any downstream benchmark, ZwZ2 improves fine-grained and spatial perception, with further gains on embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning. These results establish EvidenceConnect as an effective framework for supervising evidence connection and suggest that strengthening this shared capability leads to more generalizable multimodal perception.

1 INTRODUCTION

Spatial perception and multimodal long-context understanding depend on connecting dispersed visual evidence, but high-quality training data for these tasks are difficult and costly to build. The paper proposes EvidenceConnect and reports that training on its synthesized data improves performance across perception and downstream tasks.

1 INTRODUCTION

  • Spatial and long-context tasks rely on dispersed evidence, yet their training data are difficult to synthesize reliably and at scale.Spatial labels require precise localization and relational grounding; long-context data must support diverse questions across sequences, while embodied and agent data add environments or trajectories.Source: §1 · p. 2
  • The paper identifies connecting evidence as a shared capability: finding relevant visual facts, binding them to entities, recovering relations, and integrating them into an answer.Its steps support spatial perception and long-context understanding; identifying visual facts also supports fine-grained perception.Source: §1 · p. 2
  • EvidenceConnect synthesizes single-image questions that compare, aggregate, or chain evidence across related regions, then checks answer agreement and evidence dependence.Verified questions are converted to multiple choice and paired with the full image, so the model must locate and connect evidence itself.Source: §1 · pp. 2–3
  • ZwZ2-4B raises its 25-benchmark average from 49.41 to 54.79, with gains extending to embodied spatial reasoning, multimodal agents, and general visual reasoning.The models are trained with RLVR on synthesized data alone, without supervised fine-tuning or benchmark-specific question formats.Source: §1 · p. 3

2 RELATED WORK

The related work situates EvidenceConnect among spatial-perception, long-context, and synthetic-data approaches. It distinguishes the framework by using coherent multi-region groups and automatically checking both answer correctness and dependence on the selected regions.

Spatial perception

  • Spatial-perception work includes large-scale supervision, visual grounding, progressive training, and surrogate tasks, while EvidenceConnect targets evidence connection shared with long-context understanding.Its surrogate questions use natural images rather than targeting spatial skills alone.Source: §2 · p. 3

Long-context multimodal understanding

  • Long-context methods use multi-image or multi-turn instruction tuning and retrieval, while EvidenceConnect tests transfer from auditable single-image questions to long sequences and dialogue histories.The cited work identifies persistent difficulty aggregating distributed visual information.Source: §2 · pp. 3–4

Multimodal synthetic data

  • Figure 2 depicts the synthesis pipeline using related regions from one image, with privileged crops used for construction and the full image with outlines used for training.Its example chains a boarding-pass flight number to a departure gate and then a service icon.Source: Figure 2 · p. 4
  • EvidenceConnect builds questions from coherent groups of regions, covering aggregation and closed-set comparison as well as dependency chains, and checks answers and evidence dependence automatically.Unlike pipelines using single regions or sequential chains, its checks use cross-model agreement and region ablation without human annotation.Source: §2 · p. 4

3 BOOSTING MULTIMODAL PERCEPTION BY LEARNING TO CONNECT VISUAL EVIDENCE

EvidenceConnect constructs compositional questions over coherent groups of image regions, then verifies both answer reliability and evidence dependence. It finalizes verified questions in a policy-facing format suitable for verifiable-reward training.

3.1 MULTI-REGION QUESTION SYNTHESIS

  • Questions connect a coherent group of 3–7 regions through aggregation, closed-set comparison or matching, or dependency chains and compositions.Each region has a role, and the answer requires considering the roles together; the forms parallel spatial and long-context tasks.Source: §3.1 · p. 5
  • The synthesis process identifies related, localizable regions, groups them for a question type, and uses native-resolution crops in a layout that preserves their relative order.The question generator uses the privileged layout and region descriptions to compose candidates with concise answers and reasoning certificates.Source: §3.1 · p. 5

3.2 TWO-STAGE VERIFICATION

  • Two VLM families anchor the reference answer, while drop-one tests assess whether selected regions are necessary and reject questions with more than one redundant region.The threshold allows for noise from overlapping crop margins and variable verifier responses.Source: §3.2 · p. 6

3.3 DATA FINALIZATION

  • Verified questions are paired with a policy view of the full image, typically with region outlines, and converted to four-option items for rule-based rewards.A lightweight VLM answers each item repeatedly, and examples it always answers correctly are discarded.Source: §3.3 · p. 6

4 EXPERIMENTS

Experiments evaluate three model scales across 25 benchmarks, compare training on different synthetic datasets, and examine synthesis ablations and generalization. ZwZ2 improves across benchmarks and transfers to long-context, interactive, and other settings.

4.2 MAIN RESULTS

  • ZwZ2 improves over its base model on all 25 benchmarks at every scale, raising overall averages by 5.39, 4.96, and 4.99 points at 4B, 9B, and 27B.Table 1 reports accuracy and unweighted domain and overall averages.Source: Table 1 · p. 7, §4.2 · p. 8

4.3 COMPARISON WITH EXISTING TRAINING DATA

  • With 20.4K examples, EVIDENCECONNECT beats four synthetic datasets on all 11 benchmarks; against 73K-example ZwZ, its average is 53.86 versus 49.86 and it wins 10 of 11.It trails ZwZ only on MileBench, by 0.13 points.Source: Table 2 · p. 8, §4.3 · p. 8

4.2 MAIN RESULTS

  • Although trained only on single images, ZwZ2 improves long-context understanding at every scale, AgentVista by 4.8–6.2 points, and GUI grounding consistently.At 4B, the long-context domain average rises by 5.6 points, including a 10.0-point gain on MileBench.Source: §4.2 · p. 8
  • ZwZ2-27B achieves 60.51% overall accuracy, versus 54.02% for GPT-5.6-Luna, 57.24% for GLM-5.3-Flash, and 55.47% for DeepSeek-V4.1-Flash.It also leads the fine-grained perception domain.Source: §4.2 · p. 8

4.4 ABLATION STUDIES

  • Drop-one verification raises average accuracy from 51.03 to 52.65 points, improving 10 of 11 benchmarks, though ERQA falls by 4.00 points.Questions are rejected only when more than one region is empirically redundant.Source: Table 3 · p. 9, §4.4 · p. 9
  • Privileged-view synthesis scores 52.65 versus 51.38 average accuracy for original-image synthesis, with higher scores on 9 of 11 benchmarks.Original-image synthesis scores higher on Chartography and ERQA by 4.00 and 4.75 points, respectively.Source: Table 3 · p. 9, §4.4 · p. 9

4.5 ANALYSIS

  • Training improves average scores by 4.81 points for Qwen3-VL-2B and 2.91 points for Qwen3-VL-4B, suggesting gains extend to additional base models.The before-and-after averages are 36.82 and 41.63 for 2B, and 41.70 and 44.61 for 4B.Source: Figure 3 · p. 9, §4.5 · p. 9
  • On Qwen3.5-4B, average accuracy rises from 47.16 for the base model to 52.65 with 7K examples and 53.86 with 20.4K examples.The full training set adds 1.21 points over 7K; 7K adds 5.49 points over the base model.Source: Figure 4 · p. 9, §4.5 · p. 9

5 CONCLUSION

EVIDENCECONNECT synthesizes within-image multi-region questions, and RLVR-trained ZwZ2 models improve across 25 benchmarks at three model scales, suggesting evidence connection is transferable.

5 CONCLUSION

  • ZwZ2 improves over its base models on all 25 benchmarks at 4B, 9B, and 27B.Source: §5 · p. 10
  • EVIDENCECONNECT synthesizes multi-region visual questions and answers that require models to connect evidence within an image.Source: §5 · p. 10
  • The results suggest connecting evidence is a transferable training target, allowing easier-to-construct and verify data to benefit tasks with costly training data.Source: §5 · p. 10

A IMPLEMENTATION DETAILS

This implementation appendix details how EvidenceConnect constructs and validates multi-region training examples, covering image and region filtering, question-generation constraints, evidence-dependency checks, and conservative rewriting for unmarked images. It also summarizes prompts for evidence discovery and answer anchoring, along with distractor construction and policy-view responses.

A IMPLEMENTATION DETAILS

  • The pipeline filters SA-1B images, proposes and screens grounded regions, groups them into task-conditioned sets, and arranges native-resolution crops in a color-coded privileged layout.Images must meet resolution, entropy, edge-density, and luminance thresholds; unsuitable images or those with fewer than four retained regions are skipped. Region groups contain 3–7 regions, and layout crops include 8% margins.Source: A IMPLEMENTATION DETAILS · p. 16
  • Question candidates use task-family blueprints and explicit region contributions, then undergo certificate, answer-form, and redundancy checks before selection.The generator uses four blueprints sampled from 104 templates and can propose up to eight candidates. Rule-based checks reject invalid or disconnected reasoning certificates and candidates whose answer survives removing a region; remaining candidates are ranked and capped per group and image.Source: A IMPLEMENTATION DETAILS · pp. 16–17
  • Candidates are anchored by cross-family answer agreement, tested by adaptive drop-one ablations, and converted into marked or audited unmarked multiple-choice examples before difficulty filtering.For each selected region, the system checks whether ablation preserves the answer; examples are then screened with four responses, retaining those with one to three correct responses and capping examples at four per source image.Source: A IMPLEMENTATION DETAILS · pp. 17–18

B PROMPT TEMPLATES FOR EVIDENCECONNECT

  • The prompt-template appendix condenses templates while retaining task definitions, key constraints, and main output fields.It omits examples, formatting reminders, repeated instructions, and fields unused in this setting.Source: B PROMPT TEMPLATES FOR EVIDENCECONNECT · p. 18

B.1 EVIDENCE DISCOVERY AND QUESTION CONSTRUCTION

  • Evidence-discovery prompts request coherent, task-conditioned groups of 3–7 regions, requiring every region to contribute to an aggregate, closed set, pair, or reasoning chain.The question-realization prompt requires questions to use exactly the selected regions, with each region indispensable and the answer grounded in visible pixels; spatial questions must specify a coordinate frame.Source: §B.1 · pp. 19–20

B.2 ANSWER ANCHORING AND DEPENDENCY CHECKS

  • Answer-anchoring prompts request short, objectively supported answers and semantic-equivalence clustering, while mode-specific ablation prompts require insufficient information when a masked region prevents a complete answer.The ablation guidance covers aggregate, closed-set or pairing, and chain-dependency questions.Source: §B.2 · pp. 21–22

B.3 POLICY-VIEW AND MULTIPLE-CHOICE CONSTRUCTION

  • The unmarked-question rewrite prompt presents an anchored question and its verified answer alongside the original unmarked image.The provided passage identifies the prompt’s inputs but does not describe the remainder of the rewrite procedure.Source: §B.3 · pp. 22–23
  • Rewrites preserve the source task and answer while replacing marked-region references with unique natural descriptions or exhaustive image scope; uncertain references are flagged unsafe.An independent audit checks that targets are uniquely identifiable, the question requires multiple spatially separated regions, and every grounded evidence area is required.Source: §B.3 · p. 23
  • Distractors must be plausible, visually relevant, distinct from the answer, and matched in format and approximate length, while leaving the question stem unchanged.For open-vocabulary answers, alternatives use the same grammatical form and specificity; specified generic options are disallowed.Source: §B.3 · pp. 23–24
  • Policy-view responses answer multiple-choice questions from the marked or unmarked image using only an A–D choice in the required answer tag.The instructions prohibit option text, coordinates, grounding tokens, and markers in the final answer.Source: §B.3 · p. 24

C TRAINING DATA STATISTICS

The training set contains 20.4K examples from SA-1B images, with examples spanning three to seven regions and a broad range of difficulty levels.

C TRAINING DATA STATISTICS

  • 20.4K examples synthesized from SA-1B images make up the training set.Source: C TRAINING DATA STATISTICS · p. 24
  • Three-region examples account for 65.74% of the set and four-region examples for 24.34%, with five to seven regions making up the remainder.Source: C TRAINING DATA STATISTICS · p. 24
  • 33.75% of examples are easy, 22.04% medium, 19.07% hard, and 25.14% very hard, so more than 44% are hard or very hard.Very hard examples are retained because two stronger VLMs anchored their reference answers, despite the screening model never answering them correctly.Source: C TRAINING DATA STATISTICS · p. 24

D TRAINING DETAILS

The models are trained with GRPO for two epochs, using a rollout-based reward that checks whether the predicted option matches the ground truth.

D TRAINING DETAILS

  • All models use GRPO for two epochs, with batch size 128 and learning rate 10^-6.Source: D TRAINING DETAILS · p. 25
  • For each prompt, training samples G=8 rollouts at temperature 0.7, with a maximum response length of 32K tokens and thinking mode disabled.Source: D TRAINING DETAILS · p. 25
  • The setup uses decoupled clipping ratios ϵlow=0.2 and ϵhigh=0.3 and removes the KL penalty with βKL=0.The reward checks whether the predicted option letter matches the ground-truth option.Source: D TRAINING DETAILS · p. 25

E ATTENTION MAP COVERAGE ANALYSIS

The analysis compares ZwZ2 with its base models on annotated TreeBench evidence, measuring whether top-attended visual tokens align with annotated regions. ZwZ2 shows gains in most evaluated full-attention layers, though the metric measures alignment rather than answer correctness.

Evaluation setting

  • The evaluation uses the same 405 TreeBench examples, image–question pairs, visual-token grids, and evidence annotations for each base/trained model pair.It evaluates all eight full-attention layers at each scale.Source: E ATTENTION MAP COVERAGE ANALYSIS · p. 25

Evidence mask

  • For each example, the evidence set unions visual tokens overlapping annotated boxes, counting overlapping tokens once.All evaluated examples have at least one evidence token.Source: E ATTENTION MAP COVERAGE ANALYSIS · p. 25

Top-k evidence IoU

  • The analysis selects as many highest-weight visual tokens as there are annotated evidence tokens, then measures their spatial alignment with that evidence.It averages per-example IoU over 405 paired examples and reports differences in percentage points.Source: E ATTENTION MAP COVERAGE ANALYSIS · pp. 25–26

Aggregation and layer-wise results

  • ZwZ2 gains mean evidence IoU in 7 of 8 layers at 4B and 6 of 8 at 9B; each exception decreases by less than 0.03 percentage points.These are changes in sample means; the metric measures spatial alignment at the final prompt position, not total attention mass or answer correctness.Source: Table 5 · p. 26, E ATTENTION MAP COVERAGE ANALYSIS · p. 26

F CASE STUDY

Three selected examples contrast ZwZ2-4B with Qwen3.5-4B on cross-image category identification, question-conditioned text localization, and structured counting. Together, they illustrate integrating evidence across images, binding facts to queried references, and accounting for all instances.

Cross-image consistency

  • For the category present in all five images, ZwZ2 rejects wine glass because it is absent from two images and identifies person; the base model instead accepts majority overlap.The example highlights preserving the question’s universal scope when integrating evidence across images.Source: F CASE STUDY · p. 26, Figure 8 · p. 27

Question-conditioned text localization

  • For the black text above the lower archway, ZwZ2 identifies SEGATORI, while the base model recognizes that word but associates the question with another location.The contrast illustrates that visual text recognition must be tied to the queried location.Source: F CASE STUDY · p. 26, Figure 8 · p. 27

Structured visual counting

  • In the blister-pack example, ZwZ2 counts five tablets in each of two columns for 10 total; the base model undercounts the pack as six.The example attributes the base-model error to enumeration, not arithmetic, and shows ZwZ2 using the regular arrangement to count the complete set.Source: F CASE STUDY · pp. 26–27, Figure 8 · p. 27

Connecting evidence

  • Across the cases, the needed evidence operations are integrating facts across images, binding recognized text to a queried location, and accounting for every instance before aggregation.The examples are selected cases where ZwZ2 answers correctly and its base model fails.Source: F CASE STUDY · pp. 26–27

G LIMITATIONS AND FUTURE WORK

The authors identify limits in the supervision’s temporal scope and in the model sizes and training stages evaluated.

Limitations

  • The single-image supervision does not directly model temporal or cross-frame dependencies, leaving video-frame evidence as a natural next synthesis target.The paper reports transfer to multi-image and agent tasks despite this boundary.Source: G LIMITATIONS AND FUTURE WORK · p. 28
  • The RLVR experiments cover models up to 27B parameters; benefits for larger models or earlier training stages such as pre-training remain unstudied.Source: G LIMITATIONS AND FUTURE WORK · p. 28

Related papers

Loading 2610.37037v1…